Re: OT: high WIO on Linux

Mark Brinsmead said,on my timestamp of 10/03/2006 6:05 PM: 
Can anybody offer a plausible (even if it is hypothetical) explanation
for this behaviour?  Or maybe direct me to some resources that will help
me better understand *how* or *why* such things can happen?

I mean, really!  50% IDLE + 50% WIO is awfully strange!  My Linux skills
are maybe a bit "lightweight", but I've been working with UNIX for
decades and never seen (or at least never *noticed*) anything remotely
like this...


We had a similar situation.  But in our case it was caused
by a lot (and I mean a *lot*) of indexed io.  Random, indexed
IO.  Which if you have not enough io controllers and disk access
paths, can easily thrash the Linux 2.4 kernel IO system.

How did we fix it?  Three lines of attack:
first: increase the number of disk controllers and access paths
to the disks. We had everything going through one fc controller,
two fc controllers made a world of difference.
second: io elevators (elvtune) to very low values - typically 24.
third and best solution: run kernel 2.6 with the new io scheduler.
The wait io did not disappear - nor did we expect it to - but it
certainly went dramatically down.

Another few of things to watch out for: use aio if at all
possible.  And when doing iostat, ALWAYS do it in a cycle - never
a once of - with -d and -x <device> for each device you want and
logged on as root or equivalent.  Anything else, I've found can
subvert the results and give you a false indication of what is
really hapenning with your IO.

Again: our experience,  YMMV, etcetc.

--
Cheers
Nuno Souto
in sunny Sydney, Australia
dbvision@xxxxxxxxxxxx
--
http://www.freelists.org/webpage/oracle-l


Other related posts: